This was the question that popped into my mind a couple of days ago after I read about a Boeing 737 nose-dived into the mountains of Southern China. The past few years have been extremely rough for the entire commercial aviation sector in terms of operational safety. The fatal accidents that happened due to probable system malfunction such as MCAS have raised questions on testing and validation of aircraft systems before concluding the airworthiness of the airframe.

Much has been debated about the infamous MCAS of Boeing's 737-MAX that may have caused two fatal accidents in 2019 and 2020 killing all passengers on board. Maneuvering Characteristics Augmentation System as it is called, the MCAS system was introduced to automatically correct elevated angle of attack (AOA) by adjusting the horizontal stabilizer and the trim tab. The idea was to assist pilots to prevent a stall resulting from a steep AOA.

A modern commercial airplane is made up of thousands of complex system of systems that consists of millions of operational parameters as inputs and outputs. Some examples are numerous flow controls, temperature sensors, pressure sensors, control motors, microswitches, speed sensors, and so on. These systems and parameters have complex interdependencies with a wafer-thin margin of errors but are critical for any successful flying operation/sortie of the aircraft. Although the aviation sector has the highest safety standards and regulations that are judiciously followed by every player in the game, there is always a risk when a new system is introduced or an existing process is altered. Remember, most of the data generated in the operation of an airplane is the time series data!

Let's take the example of the MCAS system introduced into the existing aircraft system architecture. When the system was introduced, it must have been tested thoroughly in the design phase especially given it is a flight-critical system. In simple terms, depending on the AOA sensor input, MCAS activated the speed trim motor that in turn operated the jackscrew. The jackscrew then moved the horizontal stabilizer proportionate to the input. The MCAS time-series signals (AOA input, trim motor currents, jackscrew and stabilizer position, etc.) relative to other broad operational flying parameters could have been used to define AI-enabled system behaviors in a normal course of operations. The time series AI can identify relative patterns of all contributing signals for specific behavior of the system. Any deviation from this "normal" pattern could be investigated further to identify threats to usual flying operations. It could be simulated during the ground tests to detect abnormal or adverse conditions by injecting incorrect AOA input (indicating AOA failure or erroneous activation of MCAS) or trim motor failure, etc. Even the historical data from the two crashes in the form of recovered FDR could be very well analyzed to learn the precursors of such a grave disaster. The repeated activation of MCAS could definitely be detected by the time series AI software which is what possibly happened in case of real 737-MAX crashes.

The time-series AI software can be used for Operational Flight Data Monitoring (OFDM/FOQA) solutions. At a very basic level, this could be done with an offline sync and analysis of flight recorder data under aviation safety management as mandated by ICAO Annex 6 - FDA program. There has been a rapid expansion of the number of time-series data parameters that are captured using digital flight recorders or FDRs. Proactive analysis of the recorded flight data could reveal conditions that could be extremely critical to improving aviation safety. For example, erroneous and repeated activation of MCAS. The operational system data analysis will evaluate flight operations trends, identify risk precursors and novel patterns in the data, and will enable the operators to take timely remedial actions.

At a deeply integrated scale, the time-series AI can analyze an entire platform (here an aircraft) and its associated subsystems in near real-time. It can provide alerts to its operators so that an appropriate problem resolution can be identified and executed in time. This can be done by scaling the AI application with the help of state-of-the-art hardware solutions such as small embedded GPUs or advanced accelerator inferencing servers.

Finally, everything seems fine when you pose a theory but practical situations differ, especially in the case of flying machines. In my humble opinion, given the fact that it seriously affects human lives, every minute possibility of testing and validation should be explored that can potentially save disasters. Artificial intelligence and machine learning techniques have evolved tremendously in recent times and could be undoubtedly used in situations like evaluation of machine performance. It does not only saves time and effort but extensively improves quality or safety through the identification of 'hard-to-detect' problems.

References: 1. en.wikipedia.org/wiki/ 2. falkonry.com 3. datascience.aero